~/local-llm $ cat local-llm.md

Local LLM Enthusiast

DGX SparkDGX Spark RTX Pro 6000 GPURTX Pro 6000 RTX 3090 GPURTX 3090

I run local LLMs on my own GPUs. Private data stays private, latency stays low, and the bill doesn鈥檛 grow with every token. What I care about is whether a model works for a real team: how fast it runs, how it scores on evals, and whether its tool calls hold up.

Hand-drawn wizard penguin tuning a private local LLM inference engine with cache, evals, and tool output

What I work on

  • Serving models with llama.cpp, vLLM and SGLang
  • Fine-tuning open models such as Qwen, Gemma, Llama and Pi.dev variants
  • Evals with lm-evaluation-harness, evalplus and my own benchmarks
  • Tuning quantization, batching and the KV cache for throughput
  • Hooking local models up to agent tools, RAG pipelines and coding workflows

Selected work

  • Unbound and Decision Model V1, a set of Jev-style decision models, fine-tuned and released on Hugging Face.
  • DFlash support for Blackwell consumer GPUs in Luce.
  • Bonsai benchmarks on a DGX Spark (GB10).
  • A local Qwen model doing junior-dev work inside my agent workflows.

If your team needs private inference or help designing a local LLM workflow, start with my Forward Deployed Engineer page.